Liquid Types as a behavioural sandbox for agents
Summary
The article argues that current per-request or per-application permissions for AI agents are insufficient and risk exposure via the lethal trifecta. It proposes liquid types as a behavioral sandbox to enforce guardrails, demonstrated via AeonBox's sandboxed agent using linear types and uninterpreted functions to limit external access and data exfiltration. It provides code examples and discusses how these techniques could improve safety for critical systems.